Guard-Net

Contents

• Introduction
• Installation
  • Environment
  • Dependencies
• Demo
• Training
  • Data Preparation
  • SceneFlow
  • KITTI
• Testing
• Citation
• Acknowledgments

Introduction

---

The code of the paper Guard-Net: Lightweight Stereo Matching Network via Global and Uncertainty-Aware Refinement for Autonomous Driving（IEEE TITS 2024）.

Installation

---

Environment

Create a virtual environment and activate it

conda create -n guard python=3.8
source activate guard

Dependencies

install PyTorch (An example works for me)

Link：https://pan.baidu.com/s/1aUj3kB3wMkhFigTYdyHYGg?pwd=6acz

Code：6acz

pip install torchvision-0.10.1+cu111-cp38-cp38-linux_x86_64.whl
pip install torch-1.9.1+cu111-cp38-cp38-linux_x86_64.whl

install extra dependencies

pip install -r requirements.txt

Demo

---

Generate disparity images of SceneFlow test set:

Download SceneFlow Pre-trained Weights

Link：https://pan.baidu.com/s/1gbIT8CMvHTJiIMyJLFh67w?pwd=9d65

Code：9d65

python ./save_disp_sceneflow.py \
    --datapath SceneFlow Path \
    --loadckpt SceneFlow Pre-trained Weights Path

Training

---

Data Preparation

• Scene Flow Datasets

• KITTI 2012

• KITTI 2015

In our setup, the dataset is organized as follows

└── datasets
    ├── KITTI
    |   ├── 2012
    |   │   ├── training
    |   │   ├── testing
    |   |
    |   ├── 2015
    |   │   ├── training
    |   │   └── testing
    |
    └── SceneFlow
    	├── flyingthings3d__frames_finalpass
        ├── flyingthings3d__disparity
	├── driving__frames_finalpass
        ├── driving__disparity
        ├── flyingthings3d_final
        ├── monkaa__frames_finalpass
        ├── monkaa__disparity

SceneFlow

Use the following command to train Guard-Net on SceneFlow

python ./main_coex.py \
    --datapath SceneFlow Path \
    --logdir logdir Path

KITTI

Use the following command to train Guard-Net on KITTI (using pre-trained model on Scene Flow)

python ./coex_kitti.py \
    --kitti15_datapath kitti15 Path \
    --kitti12_datapath kitti12 Path \
    --logdir logdir Path \
    --loadckpt SceneFlow Pre-trained Weights Path

Testing

---

Use the following command to test Guard-Net on SceneFlow

python ./test_sceneflow.py \
    --datapath SceneFlow Path \
    --loadckpt SceneFlow Pre-trained Weights Path

Citation

---

If you find our work useful in your research, please consider citing our paper:

@article{liu2024Guard-Net,
  title={Guard-Net: Lightweight Stereo Matching Network via Global and Uncertainty-Aware Refinement for Autonomous Driving},
  author={Liu, Yujun and Zhang, Xiangchen and Luo, Yang and Hao, Qiaoqiao and Su, Jinhe and Cai, Guorong},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={25},
  number={8},
  pages={10260-10273},
  year={2024},
  publisher={IEEE}
}

Acknowledgments

---

Thanks to Antyanta Bangunharcana for open-sourcing his excellent work CoEx. Thanks to Gangwei Xu for open-sourcing his PyTorch implementation.